Habitat mapping using machine learning-extended kernel-based reclassification of an Ikonos satellite image

The spatial resolution of satellite imagery suitable for earth resources studies has improved from 80 m (Landsat-MSS, launched in 1972) to 0.6 m (QuickBird, launched in 2001). The conventional pixel-based methods developed for medium resolution satellite images are not suitable for classification of very high spatial resolution images, because the spectral responses of particular habitat classes are much more variable. On the other hand, in the original Barnsley–Barr kernel-based reclassification algorithm not only the spectral information of a pixel but also the textural information in the vicinity of the pixel is used when the pixel labeling decision is made. The first step of the kernel reclassification algorithm is to perform an initial classification of the original image. In the second step, the adjacency-event matrices are computed for each pixel according to co-occurrence frequencies of the initial classes in the kernel window. The degree of matching between an adjacency-event matrix corresponding to specific pixel and the set of class-specific template matrices produced during training is the criterion for pixel re-labeling. We extend the original kernel-based reclassification algorithm with a decision tree-based reclassification, simultaneously taking into account the class-specific similarity images, which are a side-product of the original algorithm. The advantage of decision tree-extended approach over the original approach seems to be the ability of the former to consider more input information, thus increasing the Kappa classification accuracy for an Ikonos image of our study area from 0.56 to 0.60, using a nomenclature containing 10 habitat classes.